fix(frontend): emit SGLang stream role once

[xianlubird]

Summary

• Emit delta.role only on the first choice produced by the SGLang streaming postprocessor.
• Preserve the initial assistant role when generation finishes before producing any content.
• Cover plain-text streams, finish-only streams, and reasoning-parser streams with focused unit tests.

Background

PR #12639 established the chat-stream contract at the HTTP boundary: the first emitted delta for a choice carries role: assistant, while later deltas omit the role. The SGLang Python postprocessor builds streaming choice dictionaries outside the native Rust postprocessing path, so it still attached the role to every content-bearing delta and relied on the HTTP layer to remove the repeats.

That arrangement also left an edge case at the source. When SGLang finishes immediately without generating a token, the postprocessor returns a finish-only choice with an empty delta. The HTTP deduplication logic can remove duplicate roles, but it cannot recover a role that was never emitted.

This change tracks role emission in the per-request SGLang postprocessor and attaches the role only when a choice is actually returned. Buffered UTF-8 sequences and parser-buffered output do not consume the initial role. If the first returned choice is a finish-only choice, it still carries role: assistant; any later content or terminal choices omit it.

Validation

• black --check on the modified Python files
• isort --check-only on the modified Python files
• ruff check on the modified Python files
• python3 -m py_compile on the modified Python files
• git diff --check

The targeted pytest module was not run locally because the current virtual environment does not include pytest or the SGLang test dependencies.

Summary by CodeRabbit

• Bug Fixes

  • Improved streaming responses so the assistant role appears only once, on the first response chunk.
  • Ensured finish-only responses return empty content deltas while preserving the initial assistant role.
  • Corrected role handling for fast text, reasoning, and tool-call streaming responses.

• Tests

  • Added coverage for role emission and finish-only streaming behavior.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[dynamo-ops]

/ok to test b16bd34

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[coderabbitai]

Walkthrough

Changes

SGLang role emission

Layer / File(s)	Summary
Centralized role injection components/src/dynamo/frontend/sglang_prepost.py	SglangStreamingPostProcessor tracks role emission and applies "role": "assistant" only to the first delta across fast-path and parser-based responses.
Role emission coverage components/src/dynamo/frontend/tests/test_sglang_processor_unit.py	Tests cover empty finish-only deltas, first-chunk role emission, immediate finishes, and single reasoning-stream role emission.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description explains the change and validation, but it omits the required Related Issues section and does not identify where reviewers should start.	Add the required Related Issues section and identify the specific files or code areas where reviewers should begin.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: emitting the SGLang stream role once.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

components/src/dynamo/frontend/tests/test_sglang_processor_unit.py (1)

3183-3199: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Assert that the first reasoning delta contains the role.

roles == ["assistant"] proves only that one role was emitted. It still passes if a later delta contains the role. Record emitted deltas and assert that the first returned delta has "role": "assistant" and later deltas do not. Apply the same assertion to the existing required-tool test to cover the parser path.

Suggested assertion

-        roles = []
+        emitted_deltas = []
...
-                if "role" in delta:
-                    roles.append(delta["role"])
+                emitted_deltas.append(delta)
...
-        assert roles == ["assistant"]
+        assert emitted_deltas
+        assert emitted_deltas[0]["role"] == "assistant"
+        assert all("role" not in delta for delta in emitted_deltas[1:])

The PR objective requires role emission on the first returned choice across content, finish-only, reasoning, and parser-based streams.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@components/src/dynamo/frontend/tests/test_sglang_processor_unit.py` around
lines 3183 - 3199, Update the streaming test around post.process_output to
retain each returned delta, then assert the first emitted delta contains role
"assistant" and all subsequent deltas omit role instead of only checking roles
== ["assistant"]. Apply the same first-versus-later delta assertions to the
existing required-tool test, covering both content and parser-based stream
paths.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@components/src/dynamo/frontend/tests/test_sglang_processor_unit.py`:
- Around line 3183-3199: Update the streaming test around post.process_output to
retain each returned delta, then assert the first emitted delta contains role
"assistant" and all subsequent deltas omit role instead of only checking roles
== ["assistant"]. Apply the same first-versus-later delta assertions to the
existing required-tool test, covering both content and parser-based stream
paths.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 75989a35-f2f8-48f5-b3d3-0aaa17d9c8c5

📥 Commits

Reviewing files that changed from the base of the PR and between 80865c9 and b16bd34.

📒 Files selected for processing (2)

• components/src/dynamo/frontend/sglang_prepost.py
• components/src/dynamo/frontend/tests/test_sglang_processor_unit.py

[dynamo-ops]

/ok to test 578823a

[rmccorm4]

LGTM, thanks @xianlubird